Mechanistic understanding of human–wildlife conflict through a novel application of dynamic occupancy models

Abstract

Crop and livestock depredation by wildlife is a primary driver of human–wildlife conflict, a problem that threatens the coexistence of people and wildlife globally. Understanding mechanisms that underlie depredation patterns holds the key to mitigating conflicts across time and space. However, most studies do not consider imperfect detection and reporting of conflicts, which may lead to incorrect inference regarding its spatiotemporal drivers. We applied dynamic occupancy models to elephant crop depredation data from India between 2005 and 2011 to estimate crop depredation occurrence and model its underlying dynamics as a function of spatiotemporal covariates while accounting for imperfect detection of conflicts. The probability of detecting conflicts was consistently t would not be raided in primary period t + 1 varied with elevation gradient in different seasons and was influenced negatively by mean rainfall and village density and positively by distance to forests. Negative effects of rainfall variation and distance to forests best explained variation in the probability that sites not raided by elephants in primary period t would be raided in primary period t + 1. With our novel application of occupancy models, we teased apart the spatiotemporal drivers of conflicts from factors that influence how they are observed, thereby allowing more reliable inference on mechanisms underlying observed conflict patterns. We found that factors associated with increased crop accessibility and availability (e.g., distance to forests and rainfall patterns) were key drivers of elephant crop depredation dynamics. Such an understanding is essential for rigorous prediction of future conflicts, a critical requirement for effective conflict management in the context of increasing human–wildlife interactions.

Full text

Contributed Paper
Mechanistic understanding of human–wildlife conflict
through a novel application of dynamic occupancy
models
Varun R. Goswami,∗†‡ ∗∗ Kamal Medhi,§ James D. Nichols,¶ and Madan K. Oli∗†
∗School of Natural Resources and Environment, 103 Black Hall, University of Florida, Gainesville, FL 32611, U.S.A.
†Department of Wildlife Ecology and Conservation, 110 Newins-Ziegler Hall, University of Florida, Gainesville, FL 32611, U.S.A.
‡Wildlife Conservation Society, India Program, 1669 31st Cross 16th Main, Banashankari 2nd Stage, Bengaluru 560070, India
§Samrakshan Trust, Bolsalgre, Baghmara, Meghalaya 794102, India
¶United States Geological Survey, Patuxent Wildlife Research Center, Suite 4039, 12100 Beech Forest Road, Laurel, MD 20708, U.S.A.
Abstract: Crop and livestock depredation by wildlife is a primary driver of human–wildlife conflict, a
problem that threatens the coexistence of people and wildlife globally. Understanding mechanisms that
underlie depredation patterns holds the key to mitigating conflicts across time and space. However, most
studies do not consider imperfect detection and reporting of conflicts, which may lead to incorrect inference
regarding its spatiotemporal drivers. We applied dynamic occupancy models to elephant crop depredation
data from India between 2005 and 2011 to estimate crop depredation occurrence and model its underlying
dynamics as a function of spatiotemporal covariates while accounting for imperfect detection of conflicts. The
probability of detecting conflicts was consistently <1.0 and was negatively influenced by distance to roads
and elevation gradient, averaging 0.08–0.56 across primary periods (distinct agricultural seasons within
each year). The probability of crop depredation occurrence ranged from 0.29 (SE 0.09) to 0.96 (SE 0.04).
The probability that sites raided by elephants in primary period t would not be raided in primary period
t+1 varied with elevation gradient in different seasons and was influenced negatively by mean rainfall
and village density and positively by distance to forests. Negative effects of rainfall variation and distance
to forests best explained variation in the probability that sites not raided by elephants in primary period t
would be raided in primary period t +1. With our novel application of occupancy models, we teased apart the
spatiotemporal drivers of conflicts from factors that influence how they are observed, thereby allowing more
reliable inference on mechanisms underlying observed conflict patterns. We found that factors associated with
increased crop accessibility and availability (e.g., distance to forests and rainfall patterns) were key drivers
of elephant crop depredation dynamics. Such an understanding is essential for rigorous prediction of future
conflicts, a critical requirement for effective conflict management in the context of increasing human–wildlife
interactions.
Keywords: citizen science, crop and livestock depredation, detection probability, elephants, human-dominated
landscapes, monitoring, predictive modeling
Entendimiento Mec´
anico del Conflicto Humano – Animales Silvestre a trav´
es de la Novedosa Aplicaci´
on de los
Modelos Din´
amicos de Ocupaci´
on
Resumen: La depredaci´
on de cultivos y ganado por parte de animales silvestres es un conductor principal
del conflicto humano – animales silvestres, un problema que amenaza la coexistencia de la gente y la vida
silvestre a nivel global. Entender los mecanismos que son la base de los patrones de depredaci´
on es la clave
para mitigar los conflictos a lo largo del tiempo y el espacio. Sin embargo, la mayor´
ıa de los estudios no
consideran la detecci´
on imperfecta y el reporte de conflictos, lo que puede llevar a la interferencia incorrecta
∗∗Address for correspondence: Wildlife Conservation Society, India Program, 1669 31st Cross 16th Main, Banashankari 2nd Stage, Bengaluru
560070, India. email varunr.goswami@gmail.com
Paper submitted August 8, 2014; revised manuscript accepted November 26, 2014.
1100
Conservation Biology, Volume 29, No. 4, 1100–1110
C
2015 Society for Conservation Biology.
DOI: 10.1111/cobi.12475

Goswami et al. 1101
con respecto a los conductores espacio-temporales. Aplicamos modelos din´
amicos de ocupaci´
on a datos de
depredaci´
on de cultivos por elefantes en India desde 2005 y hasta 2011 para estimar la incidencia de
depredaci´
on de cultivos y modelar sus din´
amicas como una funci´
on de covarianzas espacio-temporales
mientras representan la detecci´
on imperfecta de los conflictos. La probabilidad de detectar conflictos fue
constantemente <1.0 y estuvo influenciada negativamente por la distancia a las carreteras y el gradiente
de elevaci´
on, promediando 0.08 – 0.56 en los periodos primarios (temporadas agr´
ıcolas distintas dentro de
cada a˜
no). La probabilidad de la incidencia de depredaci´
on de cultivos vari´
o desde 0.29 (SE 0.09) hasta
0.96 (SE 0.04). La probabilidad de que los sitios saqueados por elefantes en un periodo primario t no fueran
saqueados en un periodo primario t +1vari
´
o con el gradiente de elevaci´
on en diferentes temporadas y estuvo
influenciado negativamente por la precipitaci´
on promedio y la densidad de la aldea y positivamente por la
distancia al los bosques. Los efectos negativos de la variaci´
on en la precipitaci´
on y la distancia a los bosques
explicaron de mejor manera la variaci´
on en la probabilidad de que los sitios no saqueados por elefantes
en el periodo primario t ser´
ıan saqueados en el periodo primario t +1. Con nuestra novedosa aplicaci´
on
de los modelos de ocupaci´
on, separamos a los conductores espacio-temporales de los factores que influyen
en c´
omo son observados, permitiendo as´
ılainferenciam
´
as fiable de los mecanismos que son la base de
los patrones observados de los conflictos. Encontramos que los factores asociados con el incremento en la
disponibilidad y accesibilidad de los cultivos (p. ej.: la distancia a los bosques y los patrones de precipitaci´
on)
fueron conductores clave en las din´
amicas de depredaci´
on de cultivos de los elefantes. Tal entendimiento es
esencial para una predicci´
on rigurosa de conflictos futuros, un requerimiento cr´
ıtico para el manejo efectivo
de conflictos en el contexto de las crecientes interacci´
on humano – animales silvestres.
Palabras Clave: ciencia ciudadana, depredaci´
on de cultivos y ganado, detecci´
on de probabilidad, elefantes,
modelado predictivo, monitoreo, terrenos dominados por humanos
Introduction
Conflict between people and wildlife—typically
involving species that compete with humans for space
and resources (Woodroffe et al. 2005)—is a pervasive
conservation challenge. Crop or livestock depredation
by wildlife imposes substantial costs on local people
and their livelihoods (Madhusudan 2003; Karanth et
al. 2013). Concurrently, human incursions into wildlife
habitat and retributive killing of conflict-prone species
threaten the persistence of endangered fauna living
in close proximity to people (Woodroffe & Ginsberg
1998; Goswami et al. 2014). Recurrent conflicts not only
undermine the well-being of both people and wildlife
(Madhusudan 2003), they also encumber local support
for conservation (Naughton-Treves et al. 2003).
Therefore, the effective management of human–wildlife
conflict (HWC) is an essential precondition for the
coexistence of wildlife and people across space and over
time (Madden 2004).
Crop and livestock depredation is a major source of
HWC across the world (Sukumar 2003; Treves & Karanth
2003; Woodroffe et al. 2005). Consequently, a central
focus of HWC research has been to investigate the pat-
terns and spatiotemporal correlates of crop and livestock
depredation so as to better inform conflict mitigation
strategies (e.g., Sitati et al. 2003; Gubbi 2012; Karanth
et al. 2013). Data for such studies are typically obtained
from local informants, respondents to questionnaires or
official records of reported conflicts. Within such a frame-
work, if a depredation event is reported, it is known with
certainty that it has occurred (assuming the report is
verified to avoid false positives). However, if a depre-
dation event is not recorded or reported, it could imply
either that such an event has not occurred or that it
did occur but was not detected or reported. Therefore,
the reporting of conflicts is analogous to the detection
of animals when attempting to estimate demographic
parameters or species occurrence (Williams et al. 2002;
MacKenzie et al. 2006).
Biases in conflict reporting probability may arise due to
factors such as variable search effort, whereby conflicts
from remote or inaccessible locations may be underre-
ported because of lower sampling effort in these areas.
Such concerns have recently been raised about citizen
science surveys, where imperfect detection or reporting
of data generated opportunistically under a participa-
tory research framework can potentially bias estimates
of population trends (K´
ery et al. 2010; van Strien et al.
2013). Occupancy models are designed to appropriately
account for imperfect detection and variable observation
efforts (MacKenzie et al. 2006) and can provide reliable
estimates of species occurrence in participatory research
surveys (Karanth et al. 2009; K´
ery et al. 2010). These
models could be just as effective and robust in quanti-
fying HWC when conflicts are imperfectly detected or
reported.
We used data on conflicts between people and the
Asian elephant (Elephas maximus) in Garo Hills, In-
dia, in a novel application of occupancy modeling to
HWC research. Elephants are focal species for HWC re-
search, given the damage they can inflict on human life
and livelihoods (Madhusudan 2003) and the detrimental
effects of human–elephant conflict (HEC) on elephant
Conservation Biology
Volume 29, No. 4, 2015

1102 Dynamics of Human–Wildlife Conflict
Table 1. Hypotheses and a priori predictions about the influence of spatiotemporal covariates on human–elephant conflict (HEC) detection and
reporting probabilitiesa(pt,j), as well as on the extinctiona(εt) and colonizationa(γt) of elephant crop depredation events.
Covariate effects
predicted direction of slope
Hypothesis covariate parameter (β)aestimated β(SE)b
Detection or reporting of
conflict is low in less
accessible locations.
distance to roads
ruggedness index
pt,j:−
pt,j:−
pt,j:−0.63 (0.17 ×10−3)
pt,j:−0.1 (0.03)
Elephantcropdepredation
dynamics are a function of
season-specific variation in
the spatial location of crops.
season ×ruggedness
index (RG)
εt:−for jhum (JHM) ×RG; +
for paddy (PD) ×RG and
fallow (FL) ×RG
γt:+for jhum (JHM) ×RG; −
for paddy (PD) ×RG and
fallow (FL) ×RG
εtJHM x RG:−3.84 (1.54)
εtPD x RG: 0.52 (1.85)
εtPD x RG: 2.04(3.4)
γt: no support for covariate
Elephantcropdepredation
increases with an increase in
rainfall, a key determinant of
primary productivity in
terrestrial ecosystems.
mean rainfall lagged by
2 months (Rt[−2] )
coefficient of variation of
rainfall lagged by
2 months (RCV
t[−2])
εt:−
γt:+
εt:+
γt:−
εt:−0.003 (0.001)
γt: 0.35 ×10−3(0.29 ×10−3)c
εt:−10.9 (4.41)c
γt:−6.04 (3.17)
Elephantcropdepredation
increases as accessibility to
crop fields increases.
distance to closest forest εt:+
γt:−
εt: 0.74 ×10−3(0.57 ×10−3)
γt:−0.33 (0.33 ×10−3)
village density εt:−
γt:+
εt:−2.46 (1.41)
γt: 0.77 (0.98)c
aKey:Pt,j, probability that at least one HEC event was reported during a single month j of primary period t during which at least one such event
occurred; εt, probability that sites raided by elephants in primary period t are not raided in primary period t +1; γt, probability that sites not
raided by elephants in primary period t are raided in primary period t +1.
bBased on the top model in Supporting Information.
cBased on the best supported model that included this covariate.
population persistence in areas where they co-occur with
people (Goswami et al. 2014). Depredation of cultivated
crops by elephants is the primary source of HEC in both
Africa and Asia (Sukumar 2003), and rigorous quantitative
assessments of its spatiotemporal correlates are essential
for the prediction and management of future conflicts.
Crop fields potentially represent resource-rich patches
because cultivated crops typically have higher nutritional
value and water retention capacity relative to forest veg-
etation and are more palatable (Sukumar 2003; Chiyo
et al. 2005). Therefore, depredation patterns may be ex-
pected to coincide with spatiotemporal conditions that
maximize crop availability and accessibility to elephants
(Chiyo et al. 2005; Webber et al. 2011).
We applied multiseason occupancy models (MacKen-
zie et al. 2003) to more than 6 years of elephant crop
depredation data to quantify crop depredation patterns
and to discern its potential drivers in a fragmented land-
scape while accounting for imperfect detection and re-
porting of conflicts. We tested whether accessibility of
sampling sites affected the probability of detection and
reporting of conflicts and determined whether probabili-
ties of crop depredation dynamics were driven by factors
influencing crop availability and accessibility. Specific
hypotheses and a priori predictions are in Table 1. Infer-
ences thus gathered allowed us to map potential conflict
areas and to discuss the implications of our findings for
HWC research and the management of conflicts between
elephants and people.
Methods
Study Area
Our study area in Garo Hills was a fragmented land-
scape with a mosaic of community-managed forests and
4 government-managed protected areas (PAs) (Baghmara
Reserve Forest, Balphakram National Park, Siju Wildlife
Sanctuary, and Rewak Reserve Forest) interspersed in a
matrix of agriculture and human habitation (Fig. 1). Dom-
inant agricultural land uses in the matrix included slash-
and-burn shifting cultivation (locally known as jhum),
paddy cultivation, and monoculture cash-crop planta-
tions. Road accessibility within the landscape was largely
limited to 2 major roads running in north–south and east–
west directions.
Average monthly rainfall in Garo Hills is approx-
imately 1900 mm, and most of it falls from April
to September (mean monthly rainfall during this pe-
riod is approximately 3360 mm) (Indian Institute of
Tropical Meteorology, unpublished data). Agricultural
Conservation Biology
Volume 29, No. 4, 2015

Goswami et al. 1103
Figure 1. Study area in
Garo Hills, India (black
polygon in the inset).
Protected areas included
Baghmara Reserve Forest
(BRF), Balphakram
National Park (BNP), Siju
Wildlife Sanctuary (SWS),
and Rewak Reserve Forest
(RRF). The agricultural
matrix (agriculture) was of
slash-and-burn shifting
cultivation, paddy
cultivation, and
monoculture cash-crop
plantations. Sampling sites
included grid cells of 4 km2.
seasons in the region are determined by rainfall pat-
terns and can be broadly classified as fallow sea-
son (January–March), jhum season (April–September),
and paddy season (July–December). These seasons are
defined thoroughly in the section “Analytical Design and
Occupancy Modeling.”
Quantification of Conflicts
We adapted methods that were successfully used to quan-
tify HEC in Africa (Sitati et al. 2003) for use in our study
area. We trained a team of 17 informants to record and re-
port conflicts from communal lands owned and managed
by residents of 49 villages across the study area (Fig. 1).
Informants collected data from June 2005 to October
2011. Each informant verified HEC reports from 2 to
3 villages, recorded the locations of these conflicts with
a global positioning system (GPS), and collated the infor-
mation on a standardized data collection form. We visited
each informant once a month to monitor the recording
of HEC and to retrieve the data collection forms. Crop
depredation was the primary form of HEC; reports of
property damage and human injury were negligible.
Analytical Design and Occupancy Modeling
An occupancy-modeling framework typically uses bino-
mial or multinomial data on species detection or nonde-
tection to estimate the probability of occupancy or use of
a given sampling unit (MacKenzie et al. 2006). Repeated
assessments of each sampling unit additionally allow the
estimation of species detection probability (MacKenzie
et al. 2006). We used this statistical approach to estimate
the probability of elephant crop depredation from bi-
nary conflict data (i.e., HEC reported or unreported). Our
sampling sites were 4 km2grid cells. We chose this grid
cell size because it was large enough to accommodate an
HEC event (i.e., a single HEC event is not attributed to 2
adjacent cells) but not so large as to make inferences on
the drivers of elephant crop depredation less meaningful
(Guerbois et al. 2012). We used ArcGIS (version 9.3) to
overlay a set of 43 such grid cells across the study area,
which spanned 172 km2in the agricultural matrix (Fig. 1).
We then extracted reported conflicts from within each
grid cell during the sampling period and assessed them
for detection or nondetection of HEC on a monthly basis.
Therefore, one or more reports of HEC from within a grid
cell during a month was assigned a 1, while the lack of
HEC reports in the cell was assigned a 0. In this manner,
we developed detection histories of HEC, in the form of
crop depredation, over 77 months (June 2005–October
2011) and across 43 sampled grid cells. During this pe-
riod, the study area was not sampled for 36 months, and
these months were treated as missing values (MacKenzie
et al. 2006). Similarly, grid cells that were not surveyed
during a given secondary monthly sampling occasion (de-
fined below) were treated as missing data.
We analyzed the HEC data described above with multi-
season occupancy models (MacKenzie et al. 2003, 2006).
In these models, the occupancy state of sites remains
unchanged (or changes randomly) across (kt)secondary
sampling occasions within a primary sampling period
Conservation Biology
Volume 29, No. 4, 2015

1104 Dynamics of Human–Wildlife Conflict
but may change nonrandomly among Tprimary sam-
pling periods due to colonization or local extinction.
We used seasons of crop growth and harvest to define
3 primary periods for each year of our study: fallow
season (January–April), when agriculture is limited to
household vegetable gardens; jhum crop season (May–
August), when sowing, growth, and harvest of crops
such as rice, maize, and millet occurs on hill slopes; and
paddy crop season (September–December), when wet
rice cultivation occurs in flooded valleys (Datta-Roy et al.
2009). Based on the above definition, we partitioned our
77-monthdetectionhistoryinto20primaryperiods.Each
primary period had 4 secondary monthly sampling occa-
sions (henceforth, secondary occasions), and each month
served as a temporal replicate. We expected changes in
the occurrence of crop depredation events between pri-
mary periods because of the seasonality of crop growth
and harvest. Our study included 9 primary periods with
no data, during which the grid cells were not sampled.
An example detection history is available in Supporting
Information.
For the multiseason occupancy modeling, we used pro-
gram MARK (White & Burnham 1999) implemented in R
(R Development Core Team 2013) through the RDOccu-
pEG model in the RMark library (Laake & Rexstad 2007).
We estimated the following parameters: probability of
detection and reporting of crop depredation in a grid
cell during secondary occasion jwithin primary period t
conditional on crop depredation having occurred there
(pt,j); probability of crop depredation occurrence in a
grid cell during the first primary period (ψ1); probability
that a grid cell with no crop depredation in primary pe-
riod thad crop depredation in primary period t+1(γt);
and probability that a grid cell with crop depredation in
primary period tdid not experience crop depredation
in primary period t+1(εt). The parameters γtand εt
are akin to colonization and extinction probabilities in
traditional dynamic occupancy studies (MacKenzie et al.
2003).
Our interpretation of ψwithin this framework relates
to the concept of use, defined as the occurrence of a tar-
get species—in this case, the occurrence of HEC—within
a sampling unit at random points in time (MacKenzie
et al. 2006). More specifically, we view HEC as a latent
variable, that is the potential for HEC is either present
in a location during a season or not. Our ψparameter
refers to the probability that this latent variable assumes
the value 1 (HEC potential is present). If that potential is
present, then it may be manifested (a realized HEC event
may occur), and the resulting event may or may not be
detected by our informants. The detection probability
we estimated was thus conditional on the latent HEC
variable being 1 and is the product of the probability that
an HEC event occurred and the probability of detection
of such an event given its occurrence.
We began the occupancy modeling by first identifying
the most appropriate model structure for pt,jbased
on Akaike’s information criterion corrected for small
sample sizes (AICc). Sample size was 358, calculated as
the number of surveyed grid cells times the number of
primary occasions during which they were sampled. We
used distance to major roads (defined as roads accessible
by vehicle) and elevation gradient (or ruggedness) as
covariates to test the effects of grid cell accessibility
on the probability of detection and reporting of crop
depredation events. We also allowed pt,jto vary across
primary periods. We compared these models to an
intercept-only model, where detection probability was
constant. During this analysis, we allowed εtand γtto
vary as a function of agricultural season (fallow, jhum,
or paddy), primary period, and distance to forest (i.e.,
distance of grid cell sto the closest forest). Our intention
was to use general models for εtand γtwhile identifying
the best model structure for pt,j. We modeled crop
depredation in the first primary period (ψ1) as a constant
parameter.
Next, we fixed pt,jto the best supported model struc-
tures from the previous analyses and investigated the spa-
tiotemporal drivers of εtand γt. We used the independent
and additive effects of distance to forest; village density;
agricultural season; mean rainfall during each primary
period calculated with a 2-month time lag (Rt[−2]); and
coefficient of variation (CV) of rainfall during each pri-
mary period lagged by 2 months (RCV
t[−2]). We obtained
lagged rainfall values by using rainfall estimates 2 months
prior to each month of a 4-month primary period to com-
pute rainfall means and CVs for that primary period. We
also included pairwise interactive effects of crop season
and ruggedness because agriculture in our study area is
largely practiced on hill slopes in the jhum season but in
valleys in the paddy season. We considered 2-month lags
because we expected crop productivity and availability at
a given point to be affected by prior rainfall patterns (i.e.,
during the crop growing period) and because Sukumar
(2003) reports a 2-month lag between crop growth and
peak crop depredation by elephants.
Model comparisons were made on the basis of their
AICcscores and Akaike weights (wi). We used the best
supported covariates for εtand γtto estimate predicted
probabilities of extinction and colonization of HEC in
each of the 3 crop seasons. Based on these estimates, we
derived predicted season-specific probabilities of HEC
occurrence (ψt), which we mapped across the broader
landscape to identify areas with a high potential for
conflict.
Details on how we obtained our covariate data are
provided in the Supporting Information.
Results
Seasonality of Conflicts
From 2005 to 2011, 636 crop depredation events were
reported from agricultural lands belonging to residents of
Conservation Biology
Volume 29, No. 4, 2015

Goswami et al. 1105
Figure 2. (a) Probability of detection and
reporting of crop depredation by elephants
adjacent to major roads and at minimum
ruggedness within a grid cell during each 4-month
primary period (distinct agricultural seasons
within each year). Crop seasons (fallow, jhum
[shifting cultivation], and paddy) coinciding with
the different primary periods are also indicated.
(b) Detection probability as a declining function
of distance to major roads at minimum, mean,
third quartile, and maximum ruggedness (RG).
Error bars and shading represent 95% confidence
intervals.
0.37
0.53
0.53
0.41
0.38
0.47
0.58
0.25
0.25
0.36
0.17
0.00
0.25
0.50
0.75
1.00
2005 2006 2007 2008 2009 2010 2011
Ye a r
Probability of occurrence
Season
Fallow
Jhum
Padd y
Figure 3. Probability of
occurrence of crop
depredation by elephants
during each sampled
4-month primary period
(distinct agricultural
seasons within each year)
from 2005 to 2011. Crop
seasons coinciding with
each primary period within
a year are indicated.
Missing bars coincide with
primary periods when there
was no sampling. Error
bars are 95% confidence
intervals, and numbers
above each bar represent
na¨
ıve probabilities of
conflict occurrence for each
primary period.
Conservation Biology
Volume 29, No. 4, 2015

1106 Dynamics of Human–Wildlife Conflict
the 49 surveyed villages. Annual mean number of conflict
reports for the fallow, jhum, and paddy seasons were 38,
76, and 52, respectively. Na¨
ıve probabilities of conflict
occurrence across primary periods ranged from 0.17 to
0.58 (Fig. 3).
Probability of Reporting Conflicts
The overall probability of detection and subsequent re-
porting of crop depredation per secondary sampling
occasion (pt,j) based on the best supported model where
this probability was constant was 0.37 (SE 0.03). For this
constant pt,jmodel, the estimate of detection probability
per primary occasion combined across the 4 secondary
sampling occasions (P∗) was 1 −(1 −pt,j)4=0.84. How-
ever, AICcbetween the constant model and the top
model for pt,jwas 17.61. Models that received maximum
support included an additive effect of distance to major
roads and ruggedness within a grid cell (Supporting Infor-
mation). As per the top model, estimates of pt,jfrom each
sampled primary period for grid cells that were adjacent
to major roads and had minimal ruggedness ranged from
0.21 (SE 0.07) to 0.80 (SE 0.08) (Fig. 2a). However, pt,j
was negatively influenced by both distance to major roads
and ruggedness (Table 1, Fig. 2b). We fixed to the 2 model
structures (i.e., pt,jmajor roads +ruggedness, or pt,j
major roads +ruggedness +primary period) that had
comparable support (AICc<2; Supporting Informa-
tion) (Burnham & Anderson 2002) while further model-
ing the crop depredation parameters (i.e., εt.andγt).
Spatiotemporal Patterns of Crop Depredation
Overall estimates of εtand γtbased on the intercept-
only model (i.e., a εt[.], γt[.] model), were 0.29 (SE 0.06)
and 0.34 (SE 0.05), respectively. However, there was
substantial evidence that spatiotemporal covariates af-
fected both εtand γt(Supporting Information). Pairwise
interactions between crop season and ruggedness played
an important role in accounting for the spatiotemporal
variation in εt;Akaikeweights(wi) for models that in-
cluded this effect on εtsummed to 0.82. In addition, εt
as per the top model (Supporting Information) was nega-
tively influenced by Rt[−2] and village density, whereas it
was positively influenced by distance to forest (Table 1).
The combined negative effects of RCV
t[−2] and distance to
forest best explained the variation in γt(Table 1). How-
ever, uncertainty associated with the effect of distance
to forest, particularly on γt, was substantial. Estimates of
ψtderived for the sampled primary periods based on the
top model (Supporting Information) ranged from 0.29
(SE 0.09) to 0.96 (SE 0.04) (Fig. 3).
The effects of the different spatiotemporal covariates
on εtand γt, and the resultant temporal variation in ψt,
suggested the following overall trends in elephant crop
depredation patterns. Probabilities of crop depredation
occurrence (ψt) increased as the crop season transitioned
from fallow through jhum to paddy (Fig. 3). For any
given transition of season (i.e., fallow to jhum, jhum
to paddy, or paddy to fallow) extinction probabilities
across years declined with mean rainfall lagged by 2
months (Rt[−2]) (Fig. 4). Extinction probabilities declined
marginally as village density increased and demonstrated
season-specific variation with ruggedness (Fig. 4). The
transition of season from fallow to jhum was associated
with high extinction probabilities (εt1) in grid cells
with low ruggedness irrespective of Rt[−2] or village den-
sity. In contrast, εtwas nearly 0 in cells with high rugged-
ness, although there was greater variation around these
estimates when Rt[−2] was low. As the season transitioned
from jhum to paddy, εtremained high (>0.9) in rugged
cells at Rt−21600 mm, but decreased steadily there-
after. During this period, εtin cells with low ruggedness
was negligible (<0.02) when Rt[−2] >2000 mm. Extinc-
tion probabilities did not vary with ruggedness when the
season transitioned from paddy to fallow, but it declined
when Rt[−2] and village density decreased. Colonization
probabilities declined as variability in rainfall lagged by
2 months increased (RCV
t[−2]) (Supporting Information).
As distance to forests increased, extinction probabili-
ties increased and colonization probabilities decreased,
but these effects were characterized by low precision
(Table 1).
We used the top model in Supporting Information to
predict and map potential ψfor the 3 crop seasons on
the basis of site-specific spatial covariates and season-
specific estimates of Rt[−2] and RCV
t[−2] averaged across
the sampling period between 2005 and 2011. Predicted
estimates of ψincreased as the season transitioned from
fallow through jhum to paddy and was high in grid cells
between PAs (Fig. 5).
Discussion
Occupancy Modeling of Elephant Crop Depredation
The idea of citizen science has emerged from the recog-
nition that certain types of information can only be gath-
ered through a participatory research framework (K´
ery
et al. 2010). Notwithstanding the value of such a frame-
work, concerns exist about biases arising due to imper-
fect detection or reporting of the resultant data (K´
ery
et al. 2010; van Strien et al. 2013). The quantification
of HWC largely relies on citizen science data, and fail-
ure to account for imperfect detection or reporting of
conflicts can lead to flawed inference on the patterns
and correlates of HWC. In our study, imperfect detec-
tion of a crop depredation event in a spatiotemporal
unit arose because villagers failed to detect this event or
the event was detected but went unreported. We show
how an occupancy-modeling framework can readily
Conservation Biology
Volume 29, No. 4, 2015

Goswami et al. 1107
Low VD, Low RG Low VD, High RG High VD, Low RG High VD, High RG
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
Fallow to Jhum Jhum to Paddy Paddy to Fallow
0 1000 2000 30000 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000
Mean rainfall with a 2−month time lag (mm)
Extinction probability
Figure 4. Extinction probability of crop depredation by elephants (i.e., probability that crop depredation occurred
in time t but not in time t +1) as a function of mean rainfall, calculated with a 2-month lag, during transitions of
season (right axis) for grid cells with low and high village density (VD) and ruggedness (RG). Jhum refers to
shifting cultivation. For both ruggedness and village density, low and high levels correspond to their respective
first and third quartile values: 12.99 and 16.91 ruggedness index, respectively, and 0.13 and 0.44 villages/km2
(village density), respectively. For all plots, distance to forest refuges is set to a mean value of 0.6 km.
incorporate the detection process into existing conflict-
reporting systems to account for potential biases at no
significant increase in field effort or costs.
The application of the occupancy models to our spe-
cific example system broadly suggests that the acces-
sibility of crop fields, and the availability of crops in
these fields, is a major driver of elephant crop depre-
dation patterns and dynamics. First, we found strong
season-specific signatures in elephant crop depredation
patterns. Within a year, crop availability clearly varies
among seasons, and this variation was mirrored by an
increase in probabilities of crop depredation occurrence
(ψ) from the fallow season to seasons of crop growth and
harvest (i.e., jhum and paddy seasons) (Fig. 3). Season-
and location-specific signatures in HEC extinction proba-
bilities (i.e., probabilities of sites transitioning from ex-
periencing HEC to not experiencing these problems)
best explained this seasonal variation in depredation pat-
terns (Supporting Information). The spatial locations of
crops vary by season in our study area, and this varia-
tion was evident in season-specific changes in extinction
probabilities of elephant crop depredation, particularly
when the season transitioned from fallow to jhum. Extinc-
tion probabilities were negligible in grid cells with high
ruggedness but were very high (1) in cells with low
ruggedness (Fig. 4). In contrast, the transition of season
from jhum to paddy was associated with high “extinc-
tion” probabilities in rugged cells and low “extinction”
probabilities in less rugged cells, particularly when rain-
fall levels were high (Fig. 4). Easier accessibility of crops
in the paddy season compared with those grown on hill
slopes in the jhum season likely explains the higher oc-
currence of crop depredation in the paddy season (Fig. 3).
This result could also point to a potential preference
for paddy as a food resource. Peaks in elephant crop
depredation during growing periods of preferred crops
have been reported (e.g., Sukumar 2003; Osborn 2004;
Gubbi 2012).
Conservation Biology
Volume 29, No. 4, 2015

1108 Dynamics of Human–Wildlife Conflict
Figure 5. Predicted probabilities of human–elephant conflict (HEC) occurrence across the broader
Balphakram–Baghmara landscape in the (a) fallow, (b) jhum (shifting cultivation), and (c) paddy seasons.
Second, for a given transition of season, the probabil-
ity of elephants not raiding locations that experienced
crop depredation previously, declined with an increase
in mean rainfall lagged by 2 months (Fig. 4). Furthermore,
colonization probabilities of crop depredation (i.e., prob-
abilities of sites transitioning from not being raided by
elephants to having crop depredation) decreased as the
variability in rainfall lagged by 2 months increased (Sup-
porting Information). Rainfall influences plant phenology
and periods of peak forage availability (van Shaik et al.
1993), and irregular rainfall can limit plant resource abun-
dance and introduce unpredictability in forage availability
(e.g., Knapp et al. 2002). Therefore, rainfall is conceivably
an important driver of crop productivity and availability.
Our results are consistent with findings from other sites
in Africa and Asia that suggest rainfall-related peaks in ele-
phant crop depredation patterns (Sukumar 2003; Osborn
2004; Webber et al. 2011; Gubbi 2012). The novelty of
our study lies in the insights occupancy modeling pro-
vided into how rainfall influences the dynamics that un-
derlie these patterns in elephant crop depredation while
appropriately accounting for variation in the probability
of detecting and reporting crop depredation.
Third, the estimated effects of distance to forest
refuges, albeit less precise, were positive on extinc-
tion probabilities of elephant crop depredation and neg-
ative on HEC colonization probabilities. Furthermore,
HEC extinction probabilities declined as village density
decreased. Proximity of crop fields to the forest edge
clearly provides greater crop depredation opportunity to
elephants. The presence and density of crop fields adja-
cent to forests can also be expected to increase as village
density increases in landscapes such as ours, where local
livelihoods are largely agriculture dependent. However,
all our sampled grid cells were within 2.5 km of a forest,
which is a short distance to traverse for mobile species
such as elephants. Observations from both Africa and
Asia suggest that elephant crop depredation, although
negatively affected by distance to forests, can occur up
to 4–6 km from forests occupied by elephants (Gubbi
2012; Guerbois et al. 2012). The relative proximity of
our sampled grid cells to the forest edge likely explains
the uncertainty associated with the influence of distance
to forest refuges on elephant crop depredation dynamics.
Moreover, there is an element of risk associated with crop
raiding due to human retaliation, and the perception of
such risk by elephants is hypothesized to increase with
distance from refuges (Graham et al. 2009). Therefore,
the observed positive relationship of conflict with village
density may not hold for villages far from forests.
Spatial processes beyond those induced by predictors
and conflict history of a given sampling unit, may also be
important drivers of HWC patterns. There could be neigh-
borhood effects, whereby HWC dynamics in a focal unit
are influenced by the incidence of conflict in neighboring
units. For example, the mobility of elephants might lead
Conservation Biology
Volume 29, No. 4, 2015

Goswami et al. 1109
to the prediction that a spatial unit whose neighbors ex-
perienced HEC in season tmight be expected to exhibit
a higher probability of HEC in season t+1 than a spatial
unit whose neighbors were free of conflict at t.Wedid
not consider neighborhood effects on HWC dynamics
that go beyond those imposed by chosen spatial covari-
ates. It is, however, possible to explicitly account for
the dependence of colonization and extinction rates on
neighborhood occupancy within a dynamic occupancy-
modeling framework (Yackulic et al. 2012; Eaton et al.
2014). The application of these recently developed au-
tologistic models to conflict data can provide additional
insights on the spatial processes and mechanisms that
underlie the dynamics of HWC.
Management of HWC
HWC can vary over time and space as a function of
factors such as agricultural intensity (Madhusudan
2003) and changes in human and wildlife population
densities (Treves & Karanth 2003). The occupancy-
modeling framework allowed us to directly model the
dynamics of elephant crop depredation as a function
of various spatiotemporal covariates. In so doing, we
were able to make novel inferences on the mechanisms
that underlie changes in conflict patterns over time
and across space, rather than limiting our scope of
inference to spatiotemporal variation in the patterns
themselves. This can inform the design of nuanced and
potentially more effective conflict mitigation strategies.
For example, extinction probabilities of conflict relate
to the persistence of HWC in a particular site; therefore,
where or when extinction probabilities are low, reactive
management of conflicts may be necessary. In contrast,
proactive conflict mitigation measures may be necessary
to minimize the colonization of conflict in new areas
or over time. The occupancy-modeling framework
allowed us to generate such season-specific predictions
of conflict occurrence across space (Fig. 5). Moreover,
when occupancy studies such as ours are integrated
with HWC management, predicted effects of mitigation
efforts on extinction–colonization can be tested.
Our specific findings suggest that strategies designed
to minimize elephant forays into cultivated lands may be
important for HEC mitigation. In the short term, crop-
raiding deterrents (e.g., chili pepper fences and spot-
lights) and physical barriers (e.g., electric fences) may
be effective (e.g., Davies et al. 2011). Where possible,
land-use zoning in a manner that ensures the cultivation
of depredation-prone crops (e.g., paddy and millets) away
from forested refugia may be a useful, longer term solu-
tion. Predictive maps of conflict occurrence, such as the
one we have derived for our study landscape (Fig. 5), can
be used to prioritize management by identifying high-
risk areas or can inform land-use zoning or spatial con-
servation planning (Moilanen et al. 2009). Nevertheless,
villages near forests will likely remain susceptible to ele-
phant crop depredation and other forms of HWC. It is im-
perative therefore to simultaneously undertake measures
that can potentially increase tolerance for conflict-prone
species and alleviate conflict-induced property and eco-
nomic loss. These measures could include, for instance,
encouraging affected people to report conflicts and im-
proving the processing of compensation claims (Karanth
et al. 2013) and offering communal insurance schemes to
offset economic losses due to depredation (e.g., Mishra
et al. 2003).
The prediction of future conflicts between wildlife and
people, and the design of holistic, lasting strategies that
can effectively manage these conflicts, hinges on a clear
understanding of conflict drivers over time and across
space. To that end, we demonstrate the importance of
accounting for potential biases arising from imperfect
detection or reporting of conflicts. The replicate-specific
probability of detection and reporting of elephant crop
depredation events in our study was <1.0, which sug-
gests that an assumption of perfect detectability may not
always be valid in HWC research. Explicit modeling of
the detection and reporting process is therefore critical
to tease apart the spatiotemporal correlates of HWC from
factors that drive the process by which these conflicts are
observed. For example, we found that detection proba-
bility of conflicts was particularly low for sites that were
remote and difficult to access (Fig. 2). If we had not ac-
counted for the imperfect detection of conflicts, we may
have concluded, for instance, that crop depredation is
lower in such sites—a potentially incorrect inference. We
therefore strongly recommend the adoption of sampling
designs and analytical frameworks that can account for
imperfect detection and reporting of conflicts. We are the
first to apply an occupancy-modeling framework to the
study of HWC. We demonstrate that this approach effec-
tively addresses the issue of imperfect detection, forms
a reliable and robust monitoring protocol, and allows
for inference on mechanisms underlying spatiotemporal
patterns of HWC.
Acknowledgments
We record our sincere gratitude to the U.S. Fish and
Wildlife Service – Asian Elephant Conservation Fund,
WWF – Asian Rhino and Elephant Action Strategy, In-
ternational Union for Conservation of Nature (IUCN)’s
Sir Peter Scott Fund, Ashoka Trust for Research in Ecol-
ogy and the Environment, and the University of Florida
Alumni Fellowship for financial support; Samrakshan
Trust for logistical support and for investing substantial
effort into data collection; Wildlife Conservation Society
and Centre for Wildlife Studies for institutional support;
D.I. MacKenzie and J.E. Hines for analytical advice; A.C.
Williams, A. Sharma, N. Ved, K.U. Karanth, N.S. Kumar,
and K.K. Karanth for encouragement; and R.J. Fletcher, R.
Conservation Biology
Volume 29, No. 4, 2015

1110 Dynamics of Human–Wildlife Conflict
Chellam, M.E. Sunquist, J.D. Austin, D. Vasudev, C. Ron-
dinini, G. Heard, M. K´
ery, and an anonymous reviewer
for valuable feedback that greatly helped strengthen this
paper.
Supporting Information
An example detection history (Appendix S1), details on
how we obtained our covariate data (Appendix S2),
model selection tables (Appendix S3), and spatiotem-
poral variation in conflict colonization probabilities are
available online. The authors are solely responsible for
the content and functionality of these materials. Queries
(other than absence of the material) should be directed
to the corresponding author.
Literature Cited
Burnham KP, Anderson DR. 2002. Model selection and inference: a prac-
tical information-theoretic approach. Springer-Verlag, New York.
Chiyo PI, Cochrane EP, Naughton L, Basuta GI. 2005. Temporal patterns
of crop raiding by elephants: A response to changes in forage quality
or crop availability? African Journal of Ecology 43:48–55.
Datta-Roy A, Ved N, Williams AC. 2009. Participatory elephant mon-
itoring in South Garo Hills: efficacy and utility in a human-animal
conflict scenario. Tropical Ecology 50:163–171.
Davies TE, Wilson S, Hazarika N, Chakrabarty J, Das D, Hodgson DJ,
Zimmermann A. 2011. Effectiveness of intervention methods against
crop-raiding elephants. Conservation Letters 4:346–354.
Eaton MJ, Hughes PT, Hines JE, Nichols JD. 2014. Testing metapopu-
lation concepts: effects of patch characteristics and neighborhood
occupancy on the dynamics of an endangered lagomorph. Oikos
123:662–676.
Goswami VR, Vasudev D, Oli MK. 2014. The importance of conflict-
induced mortality for conservation planning in areas of human–
elephant co-occurrence. Biological Conservation 176:191–198.
Graham MD, Douglas-Hamilton I, Adams WM, Lee PC. 2009. The move-
ment of African elephants in a human-dominated land-use mosaic.
Animal Conservation 12:445–455.
Gubbi S. 2012. Patterns and correlates of human–elephant conflict
around a south Indian reserve. Biological Conservation 148:88–95.
Guerbois C, Chapanda E, Fritz H. 2012. Combining multi-scale socio-
ecological approaches to understand the susceptibility of subsis-
tence farmers to elephant crop raiding on the edge of a protected
area. Journal of Applied Ecology 49:1149–1158.
Karanth KK, Gopalaswamy AM, Prasad PK, Dasgupta S. 2013. Patterns of
human–wildlife conflicts and compensation: insights from Western
Ghats protected areas. Biological Conservation 166:175–185.
Karanth KK, Nichols JD, Hines JE, Karanth KU, Christensen NL. 2009.
Patterns and determinants of mammal species occurrence in India.
Journal of Applied Ecology 46:1189–1200.
K´
eryM,RoyleJA,SchmidH,SchaubM,VoletB,H
¨
afliger G, Zbinden N.
2010. Site-occupancy distribution modeling to correct population-
trend estimates derived from opportunistic observations. Conserva-
tion Biology 24:1388–1397.
Knapp AK, Fay PA, Blair JM, Collins SL, Smith MD, Carlisle JD, Harper
CW, Danner BT, Lett MS, McCarron JK. 2002. Rainfall variability,
carbon cycling, and plant species diversity in a mesic grassland.
Science 298:2202–2205.
Laake JL, Rexstad EA. 2007. RMark. In Cooch EG, White
GC, editors. Program MARK: a gentle introduction. Available
from http://www.phidot.org/software/mark/docs/book (accessed
November 8, 2012).
MacKenzie DI, Nichols JD, Hines JE, Knutson MG, Franklin AB. 2003.
Estimating site occupancy, colonization, and local extinction when
a species is detected imperfectly. Ecology 84:2200–2207.
MacKenzie DI, Nichols JD, Royle JA, Pollock KH, Bailey LL, Hines JE.
2006. Occupancy estimation and modeling: inferring patterns and
dynamics of species occurrence. Elsevier, San Diego.
Madden F. 2004. Creating coexistence between humans and wildlife:
global perspectives on local efforts to address human–wildlife con-
flict. Human Dimensions of Wildlife 9:247–257.
Madhusudan MD. 2003. Living amidst large wildlife: livestock and crop
depredation by large mammals in the interior villages of Bhadra Tiger
Reserve, South India. Environmental Management 31:466–475.
Mishra C, Allen P, McCarthy T, Madhusudan MD, Bayarjargal A, Prins
HHT. 2003. The role of incentive programs in conserving the snow
leopard. Conservation Biology 17:1512–1520.
Moilanen A, Wilson KA, Possingham HP, editors. 2009. Spatial conser-
vation prioritization: quantitative methods and computational tools.
Oxford University Press, Oxford.
Naughton-Treves L, Grossberg R, Treves A. 2003. Paying for tolerance:
rural citizens’ attitudes toward wolf depredation and compensation.
Conservation Biology 17:1500–1511.
Osborn FV. 2004. Seasonal variation of feeding patterns and food se-
lection by crop-raiding elephants in Zimbabwe. African Journal of
Ecology 42:322–327.
R Development Core Team. 2013. R: A language and environment
for statistical computing. R Foundation for Statistical Computing,
Vienna, Austria.
Sitati NW, Walpole MJ, Smith RJ, Leader-Williams N. 2003. Predicting
spatial aspects of human–elephant conflict. Journal of Applied Ecol-
ogy 40:667–677.
Sukumar R. 2003. The living elephants: evolutionary ecology, behavior
and conservation. Oxford University Press, New York.
Treves A, Karanth KU. 2003. Human–carnivore conflict and perspec-
tives on carnivore management worldwide. Conservation Biology
17:1491–1499.
van Shaik CP, Terborgh JW, Wright SJ. 1993. The phenology of trop-
ical forests: adaptive significance and consequences for primary
consumers. Annual Review of Ecology and Systematics 24:353–
377.
van Strien AJ, van Swaay CAM, Termaat T. 2013. Opportunistic citizen
science data of animal species produce reliable estimates of distribu-
tion trends if analysed with occupancy models. Journal of Applied
Ecology 50:1450–1458.
Webber CE, Sereivathana T, Maltby MP, Lee PC. 2011. Elephant crop-
raiding and human-elephant conflict in Cambodia: crop selection
and seasonal timing of raids. Oryx 45:243–251.
White GC, Burnham KP. 1999. Program MARK: survival estimation from
populations of marked animals. Bird Study 46 Supplement:120–
138.
Williams BK, Nichols JD, Conroy MJ. 2002. Analysis and management
of animal populations. Academic Press, San Diego, California.
Woodroffe R, Ginsberg JR. 1998. Edge effects and the extinction of
populations inside protected areas. Science 280:2126–2128.
Woodroffe R, Thirgood S, Rabinowitz A. 2005. People and wildlife:
Conflict or coexistence? Cambridge University Press, Cambridge,
United Kingdom.
Yackulic CB, Reid J, Davis R, Hines JE, Nichols JD, Forsman E. 2012.
Neighborhood and habitat effects on vital rates: expansion of the
Barred Owl in the Oregon Coast Ranges. Ecology 93:1953–1966.
Conservation Biology
Volume 29, No. 4, 2015